load("LDS_Public_Workspace.RData")
LDSestim=LDS
table(LDSestim$Sowing_Date_Schedule)
T5_16Dec T4_15Dec T3_30Nov T2_20Nov T1_10Nov
665 1696 3167 1704 416 Spatial Exante Framework for Targeting Sowing Dates using Causal ML and Policy Learning: Eastern India Example
1 Introduction
In this notebook, I use a causal machine learning estimator, i.e., multi-armed causal random forest with augmented inverse propensity score weights (Athey et al 2019), to estimate conditional average treatment effects (CATES) for agronomic practices. These CATEs are estimated for each individual farm thereby providing personalized estimates of the potential effectiveness of the practices. I then use a debiased robust estimator in a policy tree optimization (Athey and Wager 2021) to generate optimal recommendations in the form of agronomic practices that maximize potential yield gains.
2 Preliminaries
2.1 Graphics
# Bar graphs showing percentage of farmers adopting these practices
library(tidyverse)
library(ggplot2)
bar_chart=function(dat,var){ dat|> drop_na({{var}})|> mutate({{var}}:=factor({{var}})|>fct_infreq())|> ggplot()+ geom_bar(aes(y={{var}}),fill="dodgerblue4")+ theme_minimal(base_size = 16) }
sow_plot=bar_chart(LDSestim,Sowing_Date_Schedule)+labs(y="Sowing dates")
sow_plot
library(ggpubr)
library(tidyverse)
#Sowing dates
SowingDate_Options_Errorplot= LDSestim%>%
drop_na(Sowing_Date_Schedule) %>%
ggerrorplot(x = "Sowing_Date_Schedule", y = "L.tonPerHectare",add = "mean", error.plot = "errorbar", color="steelblue", ggtheme=theme_bw())+
labs(x="Sowing date options",y="Wheat yield (t/ha)")+
theme_minimal(base_size = 16)+
coord_flip()
SowingDate_Options_Errorplot 
2.2 Descriptives
setwd("D:/OneDrive/CIMMYT/Papers/IO5.3.1.CropResponseModels/WheatResponse/Policytree_Models")
library(fBasics)
summ_stats <- fBasics::basicStats(LDSestim[,c("L.tonPerHectare","G.q5305_irrigTimes","Nperha","P2O5perha","Weedmanaged","variety_type_NMWV","Sowing_Date_Early")])
summ_stats <- as.data.frame(t(summ_stats))
# Rename some of the columns for convenience
summ_stats <- summ_stats[c("Mean", "Stdev", "Minimum", "1. Quartile", "Median", "3. Quartile", "Maximum")] %>%
rename("Lower quartile" = '1. Quartile', "Upper quartile"= "3. Quartile")
summ_stats Mean Stdev Minimum Lower quartile Median
L.tonPerHectare 2.990635 0.854990 0.2 2.4000 3.00000
G.q5305_irrigTimes 2.286183 0.767177 1.0 2.0000 2.00000
Nperha 130.220556 37.038394 0.0 105.1852 132.54321
P2O5perha 59.038865 19.629879 0.0 45.4321 59.77908
Weedmanaged 0.758990 0.427724 0.0 1.0000 1.00000
variety_type_NMWV 0.530097 0.499126 0.0 0.0000 1.00000
Sowing_Date_Early 0.277197 0.447644 0.0 0.0000 0.00000
Upper quartile Maximum
L.tonPerHectare 3.43000 6.5000
G.q5305_irrigTimes 3.00000 5.0000
Nperha 156.00000 298.4691
P2O5perha 72.69136 212.9630
Weedmanaged 1.00000 1.0000
variety_type_NMWV 1.00000 1.0000
Sowing_Date_Early 1.00000 1.00003 Causal Random Forest Model
library(grf)
library(policytree)
LDSestim_sow=subset(LDSestim, select=c("Sowing_Date_Schedule","L.tonPerHectare","I.q5505_weedSeverity_num","I.q5509_diseaseSeverity_num","I.q5506_insectSeverity_num","I.q5502_droughtSeverity_num", "Nperha","P2O5perha","variety_type_NMWV","G.q5305_irrigTimes","A.q111_fGenderdum","Weedmanaged","temp","precip","wc2.1_30s_elev","M.q708_marketDistance","nitrogen_0.5cm","sand_0.5cm", "soc_5.15cm","O.largestPlotGPS.Latitude","O.largestPlotGPS.Longitude"))
library(tidyr)
LDSestim_sow=LDSestim_sow %>% drop_na()
Y_cf_sowing=as.vector(LDSestim_sow$L.tonPerHectare)
## Causal random forest -----------------
X_cf_sowing=subset(LDSestim_sow, select=c("I.q5505_weedSeverity_num","I.q5509_diseaseSeverity_num","I.q5506_insectSeverity_num",
"Nperha","P2O5perha","variety_type_NMWV","G.q5305_irrigTimes","A.q111_fGenderdum","Weedmanaged","temp","precip","wc2.1_30s_elev",
"M.q708_marketDistance","nitrogen_0.5cm","sand_0.5cm", "soc_5.15cm","O.largestPlotGPS.Latitude","O.largestPlotGPS.Longitude"))
W_cf_sowing <- as.factor(LDSestim_sow$Sowing_Date_Schedule)
W.multi_sowing.forest <- probability_forest(X_cf_sowing, W_cf_sowing,
equalize.cluster.weights = FALSE,
seed = 2
)
W.hat.multi.all_sowing <- predict(W.multi_sowing.forest, estimate.variance = TRUE)$predictions
Y.multi_sowing.forest <- regression_forest(X_cf_sowing, Y_cf_sowing,
equalize.cluster.weights = FALSE,
seed = 2
)
print(Y.multi_sowing.forest)GRF forest object of type regression_forest
Number of trees: 2000
Number of training samples: 7562
Variable importance:
1 2 3 4 5 6 7 8 9 10 11 12 13
0.000 0.000 0.001 0.021 0.015 0.651 0.193 0.000 0.064 0.001 0.001 0.006 0.001
14 15 16 17 18
0.001 0.004 0.002 0.023 0.015 varimp.multi_sowing <- variable_importance(Y.multi_sowing.forest)
Y.hat.multi.all_sowing <- predict(Y.multi_sowing.forest, estimate.variance = TRUE)$predictions
multi_sowing.forest <- multi_arm_causal_forest(X = X_cf_sowing, Y = Y_cf_sowing, W = W_cf_sowing ,W.hat=W.hat.multi.all_sowing,Y.hat=Y.hat.multi.all_sowing,seed=2)
varimp.multi_sowing_cf <- variable_importance(multi_sowing.forest)
multi_sowing_ate=average_treatment_effect(multi_sowing.forest, method="AIPW")
multi_sowing_ate estimate std.err contrast outcome
T4_15Dec - T5_16Dec 0.2358984 0.02577961 T4_15Dec - T5_16Dec Y.1
T3_30Nov - T5_16Dec 0.4245938 0.02284938 T3_30Nov - T5_16Dec Y.1
T2_20Nov - T5_16Dec 0.5638900 0.02572197 T2_20Nov - T5_16Dec Y.1
T1_10Nov - T5_16Dec 0.6981874 0.03905789 T1_10Nov - T5_16Dec Y.1varimp.multi_sowing_cf <- variable_importance(multi_sowing.forest)
vars_sowing=c("I.q5505_weedSeverity_num","I.q5509_diseaseSeverity_num","I.q5506_insectSeverity_num",
"Nperha","P2O5perha","variety_type_NMWV","G.q5305_irrigTimes","A.q111_fGenderdum","Weedmanaged","temp","precip","wc2.1_30s_elev",
"M.q708_marketDistance","nitrogen_0.5cm","sand_0.5cm", "soc_5.15cm","O.largestPlotGPS.Latitude","O.largestPlotGPS.Longitude")
## variable importance plot ----------------------------------------------------
varimpvars_sowing=as.data.frame(cbind(varimp.multi_sowing_cf,vars_sowing))
names(varimpvars_sowing)[1]="Variableimportance_sowing"
varimpvars_sowing$Variableimportance_sowing=formatC(varimpvars_sowing$Variableimportance_sowing, digits = 2, format = "f")
varimpvars_sowing$Variableimportance_sowing=as.numeric(varimpvars_sowing$Variableimportance_sowing)
varimpplotRF_sowing=ggplot(varimpvars_sowing,aes(x=reorder(vars_sowing,Variableimportance_sowing),y=Variableimportance_sowing))+
geom_jitter(color="steelblue")+
coord_flip()+
labs(x="Variables",y="Variable importance")
previous_theme <- theme_set(theme_bw(base_size = 16))
varimpplotRF_sowing
# Policy tree --------------------------------------
DR.scores_sowing <- double_robust_scores(multi_sowing.forest)
tr_sowing <- policy_tree(X_cf_sowing, DR.scores_sowing, depth = 2)
plot(tr_sowing)tr_sowing3 <- hybrid_policy_tree(X_cf_sowing, DR.scores_sowing, depth = 3)
tr_sowing3policy_tree object
Tree depth: 3
Actions: 1: T5_16Dec 2: T4_15Dec 3: T3_30Nov 4: T2_20Nov 5: T1_10Nov
Variable splits:
(1) split_variable: O.largestPlotGPS.Latitude split_value: 25.84
(2) split_variable: P2O5perha split_value: 62.4691
(4) split_variable: O.largestPlotGPS.Latitude split_value: 25.42
(8) * action: 5
(9) * action: 4
(5) split_variable: soc_5.15cm split_value: 9.1
(10) * action: 4
(11) * action: 5
(3) split_variable: O.largestPlotGPS.Longitude split_value: 83.47
(6) split_variable: precip split_value: 710.4
(12) * action: 5
(13) * action: 4
(7) split_variable: temp split_value: 25.6
(14) * action: 2
(15) * action: 5 plot(tr_sowing3)tr_sowing4 <- hybrid_policy_tree(X_cf_sowing, DR.scores_sowing, depth = 4)
tr_sowing4policy_tree object
Tree depth: 4
Actions: 1: T5_16Dec 2: T4_15Dec 3: T3_30Nov 4: T2_20Nov 5: T1_10Nov
Variable splits:
(1) split_variable: O.largestPlotGPS.Latitude split_value: 25.84
(2) split_variable: P2O5perha split_value: 62.4691
(4) split_variable: soc_5.15cm split_value: 16.2
(8) split_variable: O.largestPlotGPS.Latitude split_value: 25.42
(16) * action: 5
(17) * action: 4
(9) split_variable: I.q5509_diseaseSeverity_num split_value: 2
(18) * action: 4
(19) * action: 5
(5) split_variable: I.q5505_weedSeverity_num split_value: 3
(10) split_variable: soc_5.15cm split_value: 9.1
(20) * action: 4
(21) * action: 5
(11) split_variable: precip split_value: 785.4
(22) * action: 4
(23) * action: 5
(3) split_variable: O.largestPlotGPS.Longitude split_value: 83.47
(6) split_variable: P2O5perha split_value: 56.7901
(12) split_variable: precip split_value: 710.4
(24) * action: 5
(25) * action: 4
(13) split_variable: temp split_value: 26.05
(26) * action: 5
(27) * action: 4
(7) split_variable: P2O5perha split_value: 74.963
(14) split_variable: temp split_value: 25.6
(28) * action: 2
(29) * action: 5
(15) split_variable: temp split_value: 26.1333
(30) * action: 5
(31) * action: 4 plot(tr_sowing4)tr_assignment_sowing=LDSestim_sow
tr_assignment_sowing$depth2 <- predict(tr_sowing, X_cf_sowing)
table(tr_assignment_sowing$depth2)
4 5
2031 5531 tr_assignment_sowing$depth3 <- predict(tr_sowing3, X_cf_sowing)
table(tr_assignment_sowing$depth3)
2 4 5
183 1252 6127 tr_assignment_sowing$depth4 <- predict(tr_sowing4, X_cf_sowing)
table(tr_assignment_sowing$depth4)
2 4 5
159 1565 5838 4 Policy learning algorithm for treatment assignment
library(rgdal)
tr_assignment_sowing$depth2_cat[tr_assignment_sowing$depth2==1]="T5_16Dec"
tr_assignment_sowing$depth2_cat[tr_assignment_sowing$depth2==2]="T4_15Dec"
tr_assignment_sowing$depth2_cat[tr_assignment_sowing$depth2==3]="T3_30Nov"
tr_assignment_sowing$depth2_cat[tr_assignment_sowing$depth2==4]="T2_20Nov"
tr_assignment_sowing$depth2_cat[tr_assignment_sowing$depth2==5]="T1_10Nov"
tr_assignment_sowingsp= SpatialPointsDataFrame(cbind(tr_assignment_sowing$O.largestPlotGPS.Longitude,tr_assignment_sowing$O.largestPlotGPS.Latitude),data=tr_assignment_sowing,proj4string=CRS("+proj=longlat +datum=WGS84"))
library(mapview)
mapviewOptions(fgb = FALSE)
tr_assignment_sowingspmapview=mapview(tr_assignment_sowingsp,zcol="depth2_cat",layer.name="Recommended sowing dates")
tr_assignment_sowingspmapview